Presentation is loading. Please wait.

Presentation is loading. Please wait.

Nonparametric Regression and Clustered/Longitudinal Data

Similar presentations


Presentation on theme: "Nonparametric Regression and Clustered/Longitudinal Data"— Presentation transcript:

1 Nonparametric Regression and Clustered/Longitudinal Data
Raymond J. Carroll Texas A&M University

2 Longitudinal nonparametric model Kernel Methods
Outline Longitudinal nonparametric model Kernel Methods Working independence pseudo observation methods Comparison With Smoothing Splines

3 Panel Data (for simplicity)
i = 1,…,n clusters/individuals j = 1,…,m observations per cluster Subject Wave 1 Wave 2 Wave m 1 X 2 n

4 The Marginal Nonparametric Model
Y = Response X = time-varying covariate Question: can we improve efficiency by accounting for correlation?

5 Independent Data Splines (smoothing, P-splines, etc.) with penalty parameter = l Ridge regression fit Some bias, smaller variance is over-parameterized least squares is a polynomial regression

6 Independent Data Kernels (local averages, local linear, etc.), with kernel density function K and bandwidth h As the bandwidth h  0, only observations with X near t get any weight in the fit

7 Kernel Methods Largely based on working independence
Ignores correlation structure entirely in the fitting Fixes up standard errors afterwards Large literature Significant loss of efficiency possible, as with any problem

8 Kernel Methods First kernel methods trying to account for correlation failed Bizarre result: Knowing the correct covariance matrix was worse than working independence Justification for working independence? Difficulty: defining “locality” for multivariate observations with the same mean function

9 Pseudo-observation Kernel Methods
Pseudo-observations transform the responses Construction: linear transformation of Y Mean = Q(X) remains unchanged Obvious(?): make the covariance matrix diagonal Apply standard kernel smoothers to independent pseudo-observations

10 Pseudo-observation Kernel Methods
Choices: infinite, but one works Note: The mean is unchanged Iterate: Start with W.I., transform, apply working independence smoother, etc. Efficiency: Always better than working independence

11 Pseudo-observation Kernel Methods
Construction: Mean = Q(X) unchanged Covariance = diagonal, back to independence Generalizes to Time Series, say AR(1) Efficiency with respect to working independence

12 Pseudo-observation Kernel Methods
Time Series: Generalizations to finite order ARMA process possible Multiple transformations chosen so that resulting estimates are asymptotically independent, then average It is not clear, however, that insisting on a transformation to independence is efficient As it turns out, in general it is not an efficient construction

13 Accounting for Correlation
Splines have an obvious analogue for non-independent data Let be a working covariance matrix Penalized Generalized least squares (GLS) GLS ridge regression Because splines are based on likelihood ideas, they generalize quickly to new problems

14 Efficiency of Splines and Pseudo-Observation Kernels: Splines Superior
Exchng: Exchangeable with correlation 0.6 AR: autoregressive with correlation 0.6 Near Sing: A nearly singular matrix

15 New Construction Due to Naisyin Wang (Biometrika, 2003) Multiple steps
Get initial estimate m observations per cluster/individual Consider observation j=1. Assume that is known and equal to for k=2,…,m Form local likelihood score with only the 1st component mean unknown

16 New Construction Continue. Consider observation j. Assume that is known and equal to for Form local likelihood score with only the jth component mean unknown Repeat for all j Sum local likelihood scores over j and solve Gives new Now iterate.

17 Efficiency of Splines and Wang-type Kernels: Nearly identical
Exchng: Exchangeable with correlation 0.6 AR: autoregressive with correlation 0.6 Near Sing: A nearly singular matrix

18 GLS Splines and New Kernels
Relationship between GLS Splines and the new kernel methods Both are pseudo-observation methods Identical pseudo-observations Working independence is applied to both pseudo-observations Fitting methods at each stage differ (splines versus kernels!) Independence?: the pseudo-observations are not

19 GLS Splines and New Kernels
Let be the inverse covariance matrix Form the pseudo-observations: Weight the jth component: Algorithm: iterate until convergence Use your favorite method (splines, kernels, etc.) This is what GLS splines and new Kernels do Not a priori obvious!

20 GLS Splines and New Kernels
It is easy to see that GLS splines have an exact formula (GLS ridge regression) Less obvious but true that the new kernel methods also have an exact formula Both are linear in the responses

21 GLS Splines and New Kernels: Locality
Write the linear expressions We generated data, fixed the first X for the first person at X11 = 0.25 Then we investigated the weight functions as a function of t, where you want to estimate the regression function

22 The weight functions WS,ij(t,X11=0. 25) and WK,ij(t,X11=0
The weight functions WS,ij(t,X11=0.25) and WK,ij(t,X11=0.25) for a specific case for correlated data, working independence Red = Kernel Blue = Spline Note the similarity of shape and the locality: only if t is near =0.25 does X11 = 0.25 get any weight

23 The weight functions WS,ij(t,X11=0. 25) and WK,ij(t,X11=0
The weight functions WS,ij(t,X11=0.25) and WK,ij(t,X11=0.25) for a specific case for correlated data, GLS Red = Kernel Blue = Spline Note the similarity of shape and the lack of locality:

24 The weight functions WS,ij(t,X11=0. 25) and WS,ij(t,X11=0
The weight functions WS,ij(t,X11=0.25) and WS,ij(t,X11=0.25) for a specific case for correlated data, GLS versus Working Independence Red = GLS Blue = Working Independence

25 Three Questions Why are neither GLS splines nor Kernels local in the usual sense? The weight functions look similar in data. Does this mean that splines and kernels are in some sense asymptotically equivalent? Theory for Kernels is possible. Can we use these results/ideas to derive bias/variance theory for GLS splines?

26 Locality GLS Splines and Kernels are iterative versions of working independence applied to Nonlocality is thus clear: if any X in a cluster or individual, say Xi1, is near t, then all X’s in that cluster, such as Xi2, get weight for Q(t) Locality is thus at the cluster/individual level

27 Spline and Kernel Equivalence
We have shown that a result similar to Silverman’s for independent data hold. Asymptotically, the spline weight function is equivalent to the same kernel weight function described by Silverman

28 Spline and Kernel Equivalence
The bandwidth though changes: for cubic smoothing splines with smoothing parameter l, let Let the density of Xij be fj Then the effective bandwidth at t is Note how this depends on the correlation structure

29 Asymptotic Theory for GLS Splines
That GLS splines have smaller asymptotic variance for same bandwidth We have derived the bias and variance formulae for cubic smoothing splines with fixed penalty parameter l 0 Without going into technical details, these formulae are the same as those for kernels with the equivalent bandwidth Generalizes work of Nychka to non-iid settings

30 Conclusions Accounting for correlation to improve efficiency in nonparametric regression is possible Pseudo-observation methods can be defined, and form an essential link GLS splines and the “right” GLS kernels have the same asymptotics Locality of estimation is at the cluster level, and not the individual Xij level.

31 Fairy Penguin Coauthors Raymond Carroll Oliver Linton Alan Welsh Series of papers summarizing these results and their history are on my web site Xihong Lin Naisyin Wang Enno Mammen

32 West Texas  East Texas  Palo Duro Canyon, the Grand Canyon of Texas
Wichita Falls, my hometown Guadalupe Mountains National Park College Station, home of Texas A&M University Midland I-45 Big Bend National Park I-35

33 Semiparametric Regression
Advertisement Semiparametric Regression Regression via penalized regression splines Cambridge University Press, 2003 David Ruppert Matt Wand Raymond Carroll


Download ppt "Nonparametric Regression and Clustered/Longitudinal Data"

Similar presentations


Ads by Google